Transforming Vision Transformer: Towards Efficient Multi-Task Asynchronous Learning